A 22-control scored assessment and the conformance spec behind it. Run it in an afternoon and find out exactly what your production agents are authorized to do — and what would stop them if something went wrong.
This is a product. You buy it, you run it yourself. It is not the Idea Stage Audit — that is where an engagement with us starts, and it is a different $499.
The gap is almost always the same thing. Your controls live in prose — a policy doc, a README, an instruction in the system prompt. Nothing refuses a violation.
A model can ignore an instruction. A feature shipped six months from now can route around an application-layer check without its author ever knowing the check existed. That is not a hypothetical: it is the documented reason a production agent platform added a database constraint only after an unrelated feature silently mis-attributed audited runs.
A model cannot ignore a CHECK constraint.
The assessment's entire job is to separate the controls you have written down
from the controls that are actually enforced.
22 controls across invocation authorization, delegation, tool access, cost containment, observability, and evidence. Each scored 0 (absent) / 1 (documented) / 2 (enforced). Produces a band from 1 to 5 and a fix list ordered by blast radius, not by score.
~45 minutes.
The implementable requirement set behind every control. MUST / SHOULD / LATER, with the actual SQL, schema shapes, and the rationale for each — including which real-world failure produced it. Every requirement maps back to an assessment control, so a failed score tells you precisely what to go build.
Six controls are hard gates. They are fail-open escalation paths — a violation grants privilege rather than merely failing to record it. Fail one and your band is capped regardless of your total. Most maturity models let a strong observability score paper over a broken authorization path. This one refuses to.
Every control traces to the Trust Plane — the layered governance model aligned to the EU AI Act, ISO 42001, and NIST AI RMF. You are not scoring against one consultant's opinion.
The constraints, the watchdog triad, the authorization/accountability split — taken from open-source agent platforms that hit the bugs first. Where a control exists because a real system shipped a privilege-escalation hole and closed it, the spec says so.
§6 lists four things nothing on the market currently solves, including tamper-evident audit records. If you need signed attestations for a regulator, the spec tells you plainly instead of selling you a checkbox.
The publicly reported agent cost blow-ups run $8,000 to $47,000. Every one of them was fan-out, not single-run cost. The control that prevents that entire class of failure is one partial unique index — about four lines of SQL, included in the spec.
If the assessment surfaces that single missing constraint in your system, the kit returns 40× to 235×. That is not a projection we invented for this page; it is control D1, and you can read the reasoning yourself.
| Option | Cost | Reality |
|---|---|---|
| Nothing | $0 | Unpriced risk. Works right up until a security review or an invoice. |
| Build the framework yourself | ~3 engineer-weeks | Read Trust Plane, the EU AI Act, ISO 42001 and NIST AI RMF, then reverse-engineer production agent platforms to find the mechanisms. |
| The Kit | $199 | An afternoon. You self-score, and you still do the remediation. |
| Kit + Findings Review | $499 | Same, plus an hour with us to check your scoring and sequence your fixes. |
| AI Governance Audit | Scoped per engagement | Three weeks, run against your actual codebase. We do the mapping and the roadmap. |
Because the assessment is a self-report, and self-reports are generous. That is not a knock on you — it is the finding this entire page opens with. Most teams score Band 2 and believe they are Band 4, and they get there one control at a time, each one scored a 2 because someone remembers writing the rule down.
The review exists to catch exactly that: the controls you scored a 2 that are actually a 1. An hour, against your real architecture, with the people who wrote the spec. If you are going to use the result to answer a customer, it is worth $300 not to have graded your own exam.
You are not buying a document. You are buying the answer to a gating question: can we deploy this for a client, pass a security review, and sign the enterprise deal? If you need that answer before you sign something, $199 is not the number you are weighing.
Instant download. Two markdown documents you can run against your system this afternoon.
Scored Band 1 or 2 with client-facing exposure? That's what the AI Governance Audit is for.
Talk to us about the audit